diff --git a/apps/device-host-agent/host_agent/app.py b/apps/device-host-agent/host_agent/app.py
index 4800c21..0ad401c 100644
--- a/apps/device-host-agent/host_agent/app.py
+++ b/apps/device-host-agent/host_agent/app.py
@@ -213,7 +213,7 @@ def create_application(
             logging.getLogger(__name__).info(
                 "MCP token generated at %s", mcp_token_path
             )
-        mcp_busy_tracker = McpBusyTracker(ttl_seconds=60.0)
+        mcp_busy_tracker = McpBusyTracker(ttl_seconds=20.0)
         executor = AssignmentExecutor(
             create_execution_factories(
                 resolved_manager,
diff --git a/apps/device-host-agent/host_agent/mcp_lock.py b/apps/device-host-agent/host_agent/mcp_lock.py
index dd33399..477ddd5 100644
--- a/apps/device-host-agent/host_agent/mcp_lock.py
+++ b/apps/device-host-agent/host_agent/mcp_lock.py
@@ -38,7 +38,7 @@ class McpBusyTracker:
     def __init__(
         self,
         *,
-        ttl_seconds: float = 60.0,
+        ttl_seconds: float = 20.0,
         now: Callable[[], datetime] | None = None,
     ) -> None:
         self._ttl = float(ttl_seconds)
diff --git a/apps/device-host-agent/host_agent/web/app.py b/apps/device-host-agent/host_agent/web/app.py
index 2ce7a18..93d2b2c 100644
--- a/apps/device-host-agent/host_agent/web/app.py
+++ b/apps/device-host-agent/host_agent/web/app.py
@@ -5,7 +5,7 @@ import base64
 import json
 from collections.abc import Awaitable, Callable
 from pathlib import Path
-from typing import Any
+from typing import TYPE_CHECKING, Any
 
 import jinja2
 from fastapi import Depends, FastAPI, HTTPException, Request
@@ -35,6 +35,9 @@ from host_agent.web.auth import (
 )
 from host_agent.web.mcp_auth import BearerAuthMiddleware
 from storage.device_config import DeviceConfigStore
+
+if TYPE_CHECKING:
+    from mcp.server.fastmcp import FastMCP
 from storage.task_metadata import TaskMetadataStore
 from storage.timeline import Timeline
 
@@ -226,7 +229,7 @@ def create_console_app(
     metadata_store: TaskMetadataStore | None = None,
     timeline: Timeline | None = None,
     executor: AssignmentExecutor | None = None,
-    mcp_server: Any = None,
+    mcp_server: FastMCP | None = None,
     mcp_token_store: McpTokenStore | None = None,
     mcp_busy_tracker: McpBusyTracker | None = None,
 ) -> FastAPI:
diff --git a/apps/device-host-agent/tests/test_app.py b/apps/device-host-agent/tests/test_app.py
index 08ec8f7..901981c 100644
--- a/apps/device-host-agent/tests/test_app.py
+++ b/apps/device-host-agent/tests/test_app.py
@@ -699,7 +699,7 @@ def test_create_application_wires_mcp_components(tmp_path, monkeypatch) -> None:
     # is mounted (responds 401, not 404) without a bearer token.
     mcp_token_store = McpTokenStore(config.identity_path.parent / "host_mcp_token.json")
     mcp_token_store.load_or_create()
-    mcp_busy_tracker = McpBusyTracker(ttl_seconds=60.0)
+    mcp_busy_tracker = McpBusyTracker(ttl_seconds=20.0)
     mcp_server = build_mcp_server(
         manager=application.heartbeat.manager,
         mcp_busy_tracker=mcp_busy_tracker,
diff --git a/apps/device-host-agent/tests/test_mcp_lock.py b/apps/device-host-agent/tests/test_mcp_lock.py
index c44ef1e..15e09d3 100644
--- a/apps/device-host-agent/tests/test_mcp_lock.py
+++ b/apps/device-host-agent/tests/test_mcp_lock.py
@@ -144,3 +144,32 @@ def test_wait_until_usable_blocks_then_fails_when_cloud_remains_busy() -> None:
     )
     assert ok is False
     assert tracker.busy_device_ids() == []
+
+
+def test_default_ttl_is_20_seconds() -> None:
+    """The McpBusyTracker default TTL is 20s: short enough to recover from a
+    dead MCP session within a heartbeat interval without an explicit release
+    callback (mcp SDK 1.28.1 has no per-session shutdown hook), but long
+    enough that an actively-busy session does not lose its lease during
+    normal operator pauses."""
+    tracker = McpBusyTracker()
+    assert tracker._ttl == 20.0
+
+
+def test_default_ttl_recovers_dead_session_within_one_window() -> None:
+    """With the 20s default, a session that never renews its lease is
+    reaped within one TTL window on the next read. This is the
+    concrete fallback behavior for I1 (no FastMCP session-end hook)."""
+    times: list[datetime] = []
+
+    def now() -> datetime:
+        return times[-1] if times else datetime(2026, 1, 1, tzinfo=UTC)
+
+    tracker = McpBusyTracker(now=now)  # default 20s TTL
+    times.append(datetime(2026, 1, 1, 12, 0, tzinfo=UTC))
+    assert tracker.acquire("phone-1", "dead-session") is True
+    # No renew: advance 21s. Lease should be swept on next read.
+    times.append(datetime(2026, 1, 1, 12, 0, 21, tzinfo=UTC))
+    assert tracker.busy_device_ids() == []
+    # New session can now acquire cleanly (no stale-busy contamination).
+    assert tracker.acquire("phone-1", "new-session") is True
diff --git a/apps/device-host-agent/tests/test_web_mcp.py b/apps/device-host-agent/tests/test_web_mcp.py
index e10f001..074fc45 100644
--- a/apps/device-host-agent/tests/test_web_mcp.py
+++ b/apps/device-host-agent/tests/test_web_mcp.py
@@ -310,3 +310,106 @@ def test_wrapped_tool_accepts_context_kwarg() -> None:
     tool = tool_manager.get_tool("take_screenshot")
     assert tool is not None
     assert tool.context_kwarg == "ctx"
+
+
+def test_busy_error_wire_shape_is_calltoolresult_iserror() -> None:
+    """Regression test for spec §7 — busy errors must be visible on the wire.
+
+    mcp SDK 1.28.1's ``Tool.run`` wraps every non-``UrlElicitationRequiredError``
+    exception (including ``McpError`` and our ``McpDeviceBusyError``) into
+    ``ToolError`` (see ``mcp/server/fastmcp/tools/base.py``). The lowlevel
+    ``call_tool`` handler then builds a ``CallToolResult(isError=True,
+    content=[TextContent(...)])`` (see
+    ``mcp/server/lowlevel/server.py::_make_error_result``). There is no public
+    path that surfaces JSON-RPC ``-32000`` + structured ``data.busy_owner`` from
+    a tool call site — the SDK's wire contract for tool errors is the
+    ``isError=true`` flag plus text content. This test pins the wire shape so
+    any future SDK upgrade that exposes a true JSON-RPC error path is caught."""
+    import asyncio
+
+    import mcp.types as types
+    from mcp.server.fastmcp import FastMCP
+    from mcp.server.fastmcp.exceptions import ToolError
+
+    manager = _make_manager_with_device()
+    tracker = McpBusyTracker()
+    status = AgentStatusTracker()
+    tracker.acquire("phone-1", "sess-other")  # different session holds the device
+
+    server: FastMCP = build_mcp_server(
+        manager=manager, mcp_busy_tracker=tracker, status_tracker=status
+    )
+    tool = server._tool_manager.get_tool("take_screenshot")  # type: ignore[attr-defined]
+    assert tool is not None
+
+    sentinel_session = object()
+    ctx = _fake_ctx(sentinel_session)
+
+    with pytest.raises(ToolError) as tool_exc:
+        asyncio.run(tool.run({"device_id": "phone-1"}, context=ctx))
+
+    # ToolError text carries the original exception message verbatim,
+    # which is what the lowlevel handler copies into TextContent.
+    message = str(tool_exc.value)
+    assert "phone-1" in message
+    assert "busy" in message
+    assert "mcp_session:sess-oth" in message  # truncated busy_owner
+
+    # The lowlevel handler converts any exception into a CallToolResult
+    # with isError=True (mcp SDK 1.28.1 — not a JSON-RPC error envelope).
+    # We invoke the SDK helper directly to lock the wire contract.
+    from mcp.server.lowlevel.server import Server as LowlevelServer
+
+    lowlevel = LowlevelServer("test-lowlevel")
+    error_result = lowlevel._make_error_result(message)  # type: ignore[attr-defined]
+    inner = error_result.root
+    assert isinstance(inner, types.CallToolResult)
+    assert inner.isError is True
+    assert len(inner.content) == 1
+    text_block = inner.content[0]
+    assert isinstance(text_block, types.TextContent)
+    assert text_block.text == message
+    # And confirm the wire shape is NOT a JSON-RPC error envelope — that
+    # would require code=-32000 + data.busy_owner, which is not exposed
+    # in mcp SDK 1.28.1 for tool-call errors.
+    assert not hasattr(inner, "code")
+    assert inner.structuredContent is None
+
+
+def test_busy_error_text_includes_cloud_assignment_owner() -> None:
+    """Same wire-shape test for the cloud_assignment branch — verifies the
+    human-readable busy_owner value (the only place to surface it given the
+    SDK forces tool errors into CallToolResult.isError=true) is correct."""
+    import asyncio
+
+    from mcp.server.fastmcp.exceptions import ToolError
+
+    manager = _make_manager_with_device()
+    tracker = McpBusyTracker()
+    status = AgentStatusTracker()
+    status.mark_assignment_started(
+        AssignmentModel(
+            task_id="t1",
+            attempt=1,
+            lease_id="l1",
+            lease_expires_at=datetime.now(UTC),
+            host_id="h1",
+            device_id="phone-1",
+            goal="cloud task",
+        )
+    )
+
+    server = build_mcp_server(
+        manager=manager, mcp_busy_tracker=tracker, status_tracker=status
+    )
+    tool = server._tool_manager.get_tool("take_screenshot")  # type: ignore[attr-defined]
+    assert tool is not None
+
+    sentinel_session = object()
+    ctx = _fake_ctx(sentinel_session)
+
+    with pytest.raises(ToolError) as tool_exc:
+        asyncio.run(tool.run({"device_id": "phone-1"}, context=ctx))
+    assert "cloud_assignment" in str(tool_exc.value)
+    assert "phone-1" in str(tool_exc.value)
+    assert "busy" in str(tool_exc.value)
diff --git a/docs/MCP_INTEGRATION.md b/docs/MCP_INTEGRATION.md
index 948d094..e566e47 100644
--- a/docs/MCP_INTEGRATION.md
+++ b/docs/MCP_INTEGRATION.md
@@ -65,8 +65,11 @@ All 11 device tools from `api/mcp.py`:
 - The cloud worker and MCP clients share the same `DeviceManager`.
 - Per-device, session-level locking: the first caller (cloud or MCP) to
   touch a device holds it; the other side sees a busy error.
-- MCP sessions hold their lock until the session ends OR 60 seconds of
-  inactivity. Cloud assignments hold theirs until the assignment
+- MCP sessions hold their lock until **20 seconds of inactivity**
+  (the `McpBusyTracker` default TTL). The mcp SDK 1.28.1 does not expose
+  a per-session shutdown callback, so a clean Hermes disconnect is also
+  recovered via the 20s TTL sweep — see the implementation note in
+  spec §6.5. Cloud assignments hold theirs until the assignment
   terminates.
 - The cloud scheduler is told about MCP-held devices via the heartbeat
   `mcp_busy_device_ids` field, so it normally won't even try to dispatch
@@ -86,13 +89,20 @@ override.
 
 ## Error responses
 
-| Condition | HTTP / JSON-RPC | Body |
+The mcp SDK 1.28.1 forces tool errors into `CallToolResult(isError=true,
+content=[TextContent(message)])` — there is no public path that surfaces
+JSON-RPC `-32000` with a structured `data.busy_owner` field from a tool
+call site. The busy-owner value lives inside the text content (full
+string for `cloud_assignment`, truncated session_id prefix for
+`mcp_session:` collisions).
+
+| Condition | JSON-RPC envelope | `result.content[0].text` |
 |---|---|---|
-| Missing/wrong bearer token | HTTP 401 | `{"error": "invalid token"}` + `WWW-Authenticate: Bearer` |
-| Device busy (cloud) | JSON-RPC `-32000` | `"device X is busy (held by cloud assignment)"`, `data.busy_owner = "cloud_assignment"` |
-| Device busy (other MCP) | JSON-RPC `-32000` | `"..."`, `data.busy_owner = "mcp_session:<prefix>"` |
-| Unknown device | JSON-RPC `-32602` | `"unknown device: X"` |
-| Tool error | JSON-RPC `-32000` | Original exception message |
+| Missing/wrong bearer token | HTTP 401 (transport-level) | `{"error": "invalid token"}` + `WWW-Authenticate: Bearer` |
+| Device busy (cloud) | `result.isError = true` | `"device <X> is busy (held by cloud assignment)"` |
+| Device busy (other MCP) | `result.isError = true` | `"device <X> is busy (held by mcp_session:<8-char-prefix>)"` |
+| Unknown device | `result.isError = false` | JSON `{"ok": false, "error": "device not found: <X>"}` |
+| Tool error | `result.isError = true` | `"Error executing tool <name>: <original-message>"` |
 
 ## Troubleshooting
 
diff --git a/docs/superpowers/specs/2026-07-21-host-agent-mcp-server-design.md b/docs/superpowers/specs/2026-07-21-host-agent-mcp-server-design.md
index fbe9065..5a707e9 100644
--- a/docs/superpowers/specs/2026-07-21-host-agent-mcp-server-design.md
+++ b/docs/superpowers/specs/2026-07-21-host-agent-mcp-server-design.md
@@ -38,7 +38,7 @@
 | D6 | 设备状态映射 = 走 `_device_display_status()` 同款逻辑 | 避免"所有连上的设备看起来都 busy" |
 | D7 | Cloud worker 与 MCP server 在同一 host-agent 进程并存 | 不互斥，共享 `DeviceManager` |
 | D8 | Cloud ↔ MCP 协调 = 心跳上报 `mcp_busy_device_ids`，cloud scheduler 跳过 | 心跳 schema 扩展，cloud 侧 _matches() 一处改动 |
-| D9 | MCP session 级 lazy acquire 锁，60s TTL 兜底 | session_id 来自 FastMCP 上下文 |
+| D9 | MCP session 级 lazy acquire 锁，20s TTL 兜底（mcp SDK 1.28.1 无 session-end callback，见 §6.5） | session_id 来自 FastMCP 上下文 |
 | D10 | Skill catalog 工具 MVP 不暴露，保留 `create_mcp_server(skill_catalog_store=...)` 参数化挂载点 | 未来可加 mutating 工具 |
 | D11 | `wait_until_usable` 方法实现 + 单测，但调用方不接入 | 预留能力，MVP 全部 fail-fast |
 | D12 | `tool_handlers(manager)` 改为必传 | 已 grep 确认无调用方依赖 None 默认，根治 `DeviceNotFoundError` 类静默回退地雷 |
@@ -102,7 +102,7 @@ class McpDeviceLease:
     last_seen_at: datetime
 
 class McpBusyTracker:
-    def __init__(self, *, ttl_seconds: float = 60.0, now=None) -> None: ...
+    def __init__(self, *, ttl_seconds: float = 20.0, now=None) -> None: ...
     def acquire(self, device_id: str, session_id: str) -> bool: ...
     def renew(self, device_id: str, session_id: str) -> bool: ...
     def release(self, session_id: str) -> list[str]: ...
@@ -258,32 +258,44 @@ run_async() 主循环启动（行为不变）：
 ### 6.5 Session 结束 / TTL 过期
 
 ```
-正常：Hermes 主动断开 → FastMCP 触发 session shutdown callback
-      → mcp_busy_tracker.release(session_id) → 该 session 持有的所有 lease 释放
-      → 下次心跳 payload 不再包含这些 device_id → cloud 重新视为 idle
+正常：Hermes 主动断开 → mcp SDK 1.28.1 没有 per-session shutdown hook
+      → 该 session 的 lease 进入 TTL 倒计时
+      → 20s TTL 到期 → 下次 busy_device_ids() 或 snapshot() 调用时 lazy sweep
+      → lease 清理 → 下次心跳 payload 不再包含 → cloud 重新视为 idle
 
-异常：Hermes 崩溃 / 网络断 → 无 shutdown callback
-      → 60s TTL 到期 → 下次 busy_device_ids() 或 snapshot() 调用时 lazy sweep
+异常：Hermes 崩溃 / 网络断 → 同上，无 shutdown callback
+      → 20s TTL 到期 → 下次 busy_device_ids() 或 snapshot() 调用时 lazy sweep
       → lease 清理 → 下次心跳 payload 不再包含 → cloud 重新视为 idle
 ```
 
+**Implementation note (2026-07-21 fix wave):** mcp SDK 1.28.1 exposes
+only a server-level `lifespan` hook; `ServerSession.__aexit__` and
+`StreamableHTTPSessionManager` do not surface a per-session
+shutdown callback. The spec originally described a release-on-clean-
+disconnect path that the SDK cannot deliver today. The fallback is
+the 20-second TTL sweep — short enough that a normal heartbeat
+interval (30s) catches the recovery before the cloud scheduler
+notices, long enough that an actively-busy session does not lose its
+lease during normal operator pauses. Explicit release on session end
+remains a future enhancement if/when the SDK exposes the hook.
+
 ## 7. Error Handling Matrix
 
 | # | 触发条件 | 返回语义 | 备注 |
 |---|---|---|---|
 | 1 | `Authorization` 缺失/不匹配 | HTTP 401 + `WWW-Authenticate: Bearer` + JSON `{"error":"invalid token"}` | 不写失败日志；首次成功鉴权写 INFO |
-| 2 | Cloud 占用目标设备 | JSON-RPC `-32000` + `"device X is busy (held by cloud assignment)"` + `data.busy_owner = "cloud_assignment"` | DEBUG 日志 |
-| 3 | 另一 MCP session 占用 | JSON-RPC `-32000` + `"device X is busy (held by another MCP session)"` + `data.busy_owner = "mcp_session:<prefix>"` | DEBUG 日志 |
-| 4 | `device_id` 不存在 | JSON-RPC `-32602` + `"unknown device: X"` | 复用 `call_with_semantic_errors` |
-| 5 | 工具底层异常 | JSON-RPC `-32000` + 原异常 message | WARNING + exc_info |
+| 2 | Cloud 占用目标设备 | `CallToolResult(isError=true, content=[TextContent("device phone-1 is busy (held by cloud assignment)")])` | DEBUG 日志 |
+| 3 | 另一 MCP session 占用 | `CallToolResult(isError=true, content=[TextContent("device phone-1 is busy (held by mcp_session:<8-char-prefix>)")])` | DEBUG 日志 |
+| 4 | `device_id` 不存在 | `CallToolResult(isError=false)` + JSON `{"ok": false, "error": "device not found: phone-1"}` | 复用 `call_with_semantic_errors`；语义错误不抛 |
+| 5 | 工具底层异常 | `CallToolResult(isError=true, content=[TextContent("Error executing tool <name>: <orig-msg>")])` | WARNING + exc_info |
 | 6 | Cloud assignment 启动前命中 MCP 占用 | `AssignmentExecutionResult(status="failed", failure_reason="device held by active MCP session")` | INFO 一次 |
 | 7 | Token 文件损坏 JSON | host-agent 启动失败，stderr 提示 | 不静默重新生成 |
 | 8 | Token 文件不可写 | host-agent 启动失败 | 同上 |
-| 9 | MCP session 异常断开 | lease 进入 TTL 倒计时 | 60s 后 lazy sweep |
+| 9 | MCP session 异常断开 | lease 进入 TTL 倒计时 | 20s 后 lazy sweep |
 | 10 | TTL 过期瞬间 Hermes 重连 | renew 容忍边界：session_id 匹配 → 重新 acquire 而非报错 | 无感 |
 | 11 | 同 session 并发不同设备 | 各自独立 acquire | per-device 设计 |
 | 12 | 同 session 并发同一设备 | 第一个 acquire；第二个 renew（同 session_id） | 并发 safe |
-| 13 | FastMCP 提取不到 session_id | JSON-RPC `-32001` + `"cannot determine MCP session"` | ERROR + exc_info |
+| 13 | FastMCP 提取不到 session_id | `CallToolResult(isError=true, content=[TextContent("cannot determine MCP session")])` | ERROR + exc_info |
 | 14 | host-agent 关停时有 active MCP session | lease 随进程退出消失 | 不需显式清理 |
 
 ### 错误返回格式
@@ -292,17 +304,32 @@ run_async() 主循环启动（行为不变）：
 {
   "jsonrpc": "2.0",
   "id": "<request-id>",
-  "error": {
-    "code": -32000,
-    "message": "device phone-1 is busy (held by cloud assignment)",
-    "data": {
-      "device_id": "phone-1",
-      "busy_owner": "cloud_assignment"
-    }
+  "result": {
+    "content": [
+      {
+        "type": "text",
+        "text": "device phone-1 is busy (held by cloud assignment)"
+      }
+    ],
+    "isError": true
   }
 }
 ```
 
+**Implementation note (2026-07-21 fix wave):** mcp SDK 1.28.1's
+`Tool.run` wraps every non-`UrlElicitationRequiredError` exception
+(including `McpError` with a typed `ErrorData`) into `ToolError`.
+The lowlevel `call_tool` handler then serializes any exception as
+`CallToolResult(isError=true, content=[TextContent(message)])` via
+`_make_error_result`. There is no public path that surfaces JSON-RPC
+`-32000` with a structured `data.busy_owner` field from a tool call
+site — the SDK's wire contract for tool errors is the `isError=true`
+flag plus text content. The busy-owner value lives in the text
+content (truncated session_id for `mcp_session:` collisions, full
+string for `cloud_assignment`). Unknown-device and other semantic
+errors are returned as normal `ok=False` payloads inside a successful
+`CallToolResult(isError=false)` (see `api/mcp.py::call_with_semantic_errors`).
+
 ## 8. Testing Strategy
 
 ### 8.1 单元测试
@@ -376,13 +403,13 @@ run_async() 主循环启动（行为不变）：
 - **Q2**：`mcp-token` CLI 是否需要鉴权？→ MVP 不鉴权，假定能访问宿主机的操作者可信（与 `setup` 子命令同款）。后续可加 `--password` 校验 local_account。
 - **Q3**：MCP 工具调用是否需要 wall-clock 超时？→ MVP 不加，依赖底层超时；如有"Hermes 调用挂死"报告再加。
 - **Q4**：心跳扩展是 Alembic migration 还是仅 schema 字段？→ MVP 仅 transient 字段（heartbeat 接收 → scheduler 用 → 丢弃），无需 migration。
-- **Q5（待实测）**：60s TTL 是否合适？→ 实测后调整。太短容易误释放正常 session 的锁；太长 Hermes 崩溃后设备不可用窗口大。
+- **Q5（已答，2026-07-21 fix wave）**：TTL 从 60s 降到 20s，因为 mcp SDK 1.28.1 没有 per-session shutdown callback（见 §6.5）。20s 仍能容许正常 operator 暂停，但能在一次 30s 心跳窗口内回收崩盘 session 的锁；新 `test_default_ttl_is_20_seconds` 和 `test_default_ttl_recovers_dead_session_within_one_window` 锁定该值。如有"Hermes 长操作横跨 20s 静默"报告再调高。
 
 ## 11. Risks
 
 - **R1**：Cloud 心跳窗口期（30s）冲突。缓解：AssignmentExecutor 启动前 fail-fast。残留：cloud 可能基于过期心跳派任务、host-agent fail、cloud 重试——浪费 attempt 配额。**接受**。
 - **R2**：Hermes 长时间占用设备导致 cloud 任务反复 fail。MVP 无自动缓解；用户手动管控；未来启用 `wait_until_usable`。**接受**。
-- **R3**：Hermes 崩溃后 60s 内设备不可用。TTL 兜底但窗口存在。**接受，记入 Q5**。
+- **R3**：Hermes 崩溃后 ≤20s 设备不可用（TTL 兜底，但窗口存在）。窗口仍 ≤30s 心跳间隔，cloud 侧下次心跳能学到。**接受，记入 Q5**。
 - **R4（已消除）**：`tool_handlers` 签名改动影响面。Grep 确认 0 调用方依赖 None 默认。
 - **R5**：FastMCP `session_id` 提取依赖 `mcp` SDK 内部 API。缓解：e2e 测试覆盖；SDK 升级 CI 能及时暴露。
 - **R6**：`mcp` SDK 需作为 device-host-agent 直接依赖（目前通过 Runtime 传递）。需加进 `apps/device-host-agent/pyproject.toml`（与 `filelock` 直接化先例一致）。